-
Notifications
You must be signed in to change notification settings - Fork 0
/
bert.py
219 lines (155 loc) · 9.1 KB
/
bert.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
import sentencepiece
import tensorflow as tf
from tensorflow.keras import layers, models
class PretrainerBERTv2(models.Model):
def __init__(self, num_layers, vocab_size, seq_len, hidden_size, dff, num_heads, dropout_rate=0.1):
super().__init__()
self.num_layers = num_layers
self.vocab_size = vocab_size
self.seq_len = seq_len
self.hidden_size = hidden_size
self.dff = dff
self.num_heads = num_heads
self.bert = BERT(self.num_layers, self.vocab_size, self.seq_len, self.hidden_size, self.dff, self.num_heads, dropout_rate)
self.dense_for_nsp = layers.Dense(1, activation='sigmoid', kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.pool = layers.Dense(self.hidden_size, activation='tanh', kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.dense_for_mlm = layers.Dense(self.vocab_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.dropout = layers.Dropout(dropout_rate)
self.layernorm = layers.LayerNormalization()
self.dense = layers.Dense(self.hidden_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02), activation='gelu')
def call(self, input_ids, seg_ids, mask, training=True):
x = self.bert(input_ids, seg_ids, mask, training)
pooled_output = self.pool(x[:, 0])
nsp_prediction = self.dense_for_nsp(pooled_output)
mlm_prediction = self.dense(x)
mlm_prediction = self.dropout(mlm_prediction, training=training)
mlm_prediction = self.layernorm(mlm_prediction)
mlm_prediction = self.dense_for_mlm(mlm_prediction)
return mlm_prediction, nsp_prediction, pooled_output, x
class PretrainerBERT(models.Model):
def __init__(self, num_layers, vocab_size, seq_len, hidden_size, dff, num_heads, dropout_rate=0.1):
super().__init__()
self.num_layers = num_layers
self.vocab_size = vocab_size
self.seq_len = seq_len
self.hidden_size = hidden_size
self.dff = dff
self.num_heads = num_heads
self.bert = BERT(self.num_layers, self.vocab_size, self.seq_len, self.hidden_size, self.dff, self.num_heads, dropout_rate)
self.dense_for_nsp = layers.Dense(1, activation='sigmoid', kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.dense_for_mlm = layers.Dense(self.vocab_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
def call(self, input_ids, seg_ids, mask, training=True):
x = self.bert(input_ids, seg_ids, mask, training)
nsp_prediction = self.dense_for_nsp(x[:,0])
mlm_prediction = self.dense_for_mlm(x)
return mlm_prediction, nsp_prediction, x[:,0], x
class BERT(layers.Layer):
def __init__(self, num_layers, vocab_size, seq_len, hidden_size, dff, num_heads, dropout_rate=0.1):
super().__init__()
self.num_layers = num_layers
self.vocab_size = vocab_size
self.seq_len = seq_len
self.hidden_size = hidden_size
self.dff = dff
self.num_heads = num_heads
self.embedding_layer = EmbeddingProcessor(self.vocab_size, self.seq_len, self.hidden_size, dropout_rate)
self.transformers = Transformer(self.num_layers, self.hidden_size, self.dff, self.num_heads, dropout_rate)
def call(self, input_ids, seg_ids, attn_mask, training=True):
x = self.embedding_layer(input_ids, seg_ids, training)
x = self.transformers(x, attn_mask, training)
return x
class EmbeddingProcessor(layers.Layer):
def __init__(self, vocab_size, seq_len, hidden_size, dropout_rate=0.1, initialize_range=0.02):
super().__init__()
self.vocab_size = vocab_size
self.seq_len = seq_len
self.hidden_size = hidden_size
self.seq_list = tf.reshape(tf.range(self.seq_len), [1, -1])
self.position_embedding = layers.Embedding(input_dim=self.seq_len,
output_dim=self.hidden_size,
embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=initialize_range))
self.segment_embedding = layers.Embedding(input_dim=2,
output_dim=self.hidden_size,
embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=initialize_range))
self.embedding = layers.Embedding(input_dim=self.vocab_size,
output_dim=self.hidden_size,
embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=initialize_range))
self.layer_norm = layers.LayerNormalization()
self.dropout = layers.Dropout(dropout_rate)
def call(self, input_ids, seg_ids, training=True):
embedding = self.embedding(input_ids)
embedding += self.position_embedding(self.seq_list)
embedding += self.segment_embedding(seg_ids)
embedding = self.layer_norm(embedding)
embedding = self.dropout(embedding, training=training)
return embedding
class Transformer(layers.Layer):
def __init__(self, num_layers, hidden_size, dff, num_heads, dropout_rate=0.1):
super().__init__()
self.num_layers = num_layers
self.hidden_size = hidden_size
self.num_heads = num_heads
self.encoders = [TransformerBlock(self.hidden_size, dff, self.num_heads, dropout_rate) for _ in range(num_layers)]
def call(self, x, att_mask, training=True):
for i in range(self.num_layers):
x = self.encoders[i](x, att_mask, training)
return x
class TransformerBlock(layers.Layer):
def __init__(self, hidden_size, dff, num_heads, dropout_rate=0.1):
super().__init__()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.dff = dff
self.attention_layer = AttentionLayer(self.hidden_size, self.num_heads, dropout_rate)
self.projection_layer_1 = ProjectionLayer(self.hidden_size, dropout_rate)
self.projection_layer_2 = ProjectionLayer(self.hidden_size, dropout_rate)
self.point_wise_feed_forward = layers.Dense(self.dff, activation='gelu', kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
def call(self, x, att_mask, training=True):
attention_output = self.attention_layer(x, att_mask, training)
x = self.projection_layer_1(attention_output, x, training)
intermidiate = self.point_wise_feed_forward(x)
x = self.projection_layer_2(intermidiate, x, training)
return x
class AttentionLayer(layers.Layer):
def __init__(self, hidden_size, num_heads, dropout_rate=0.1):
super().__init__()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.depth = self.hidden_size // self.num_heads
self.wq = layers.Dense(hidden_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.wk = layers.Dense(hidden_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.wv = layers.Dense(hidden_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.dropout = layers.Dropout(dropout_rate)
def call(self, x, att_mask, training=True):
batch_size = tf.shape(x)[0]
q = self.wq(x)
k = self.wk(x)
v = self.wv(x)
q = tf.transpose(tf.reshape(q, [batch_size, -1, self.num_heads, self.depth]), perm=[0, 2, 1, 3])
k = tf.transpose(tf.reshape(k, [batch_size, -1, self.num_heads, self.depth]), perm=[0, 2, 1, 3])
v = tf.transpose(tf.reshape(v, [batch_size, -1, self.num_heads, self.depth]), perm=[0, 2, 1, 3])
attention_scores = tf.einsum('bnqd,bnkd->bnqk', q, k)
attention_scores = attention_scores / tf.sqrt(float(self.depth))
# {1, 0} -> {0.0, -inf}
att_mask = tf.expand_dims(tf.expand_dims(att_mask, 1), 1) * -10000.
attention_scores = tf.add(attention_scores, att_mask)
# [b, n, q, k]
attention_weights = tf.nn.softmax(attention_scores, axis=-1)
attention_weights = self.dropout(attention_weights, training=training)
output = tf.matmul(attention_weights, v)
output = tf.transpose(output, perm=[0, 2, 1, 3])
output = tf.reshape(output, shape=[batch_size, -1, self.hidden_size])
return output
class ProjectionLayer(layers.Layer):
def __init__(self, hidden_size, dropout_rate=0.1):
super().__init__()
self.hidden_size = hidden_size
self.dropout_rate = dropout_rate
self.dense = layers.Dense(self.hidden_size, kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
self.dropout = layers.Dropout(self.dropout_rate)
self.layer_norm = layers.LayerNormalization()
def call(self, output, residual, training=True):
output = self.dense(output)
output = self.dropout(output, training=training)
output = self.layer_norm(output + residual)
return output